Projects

FastTrees model

FastTrees model

FastTrees is a general purpose neural module for fast sequence encoding which imbues hierarchical inductive biases to improve Transformers and LSTM. It outperforms existing state-of-the-art models on logical inference tasks by +4% and mathematical language understanding by +8%

Paper

Parse Tree

Orchard dataset

Orchard is a new benchmark dataset for measuring systematic generalization of multi-hierarchical reasoning, which is a key task in modelling languages, which are recursive in nature. Backed by a set of rigorous experiments, we show that state-of-the-art NLP models of Transformer and LSTM surprisingly fail in this systematic generalization task.

Paper

Tweet Classsification

Size v. Specificity

Comparing pre-trained BERT, GloVe and learned embeddings for Disaster Tweet Classification, we examined the trade-off between larger pre-trained corpus size with matching levels between source and target domain for transfer learning.

Github

SIH

Singapore-India Hackathon 2019

DrishyaAI is a dashboard for policy makers that uses neural networks (LSTM, GRU, RNN) to predict future disease hotspots, and optimization solvers to solve supply chain crunch of healthcare supplies.

Github